MDX-Tom/gpt-5.6-instruct

The `gpt-5.6-instruct` repository by MDX-Tom provides a jailbreak/prompt-injection toolkit ("破甲提示词") targeting the `gpt-5.6-sol` model (OpenAI Codex). Its core purpose is to reframe security research, penetration testing, reverse engineering, software cracking, and NSFW fictional content as local sandbox tasks, suppressing refusal behaviors and improving direct instruction execution. The project's standout feature is a self-iterating optimization pipeline: the model itself absorbs user instructions, real failure cases, and GitHub Issues to update test sets, analyze regressions, and rewrite prompts — with a three-tier (low/medium/high) regression gate controlling releases. The current default version, `v41`, achieves 120/120 on the original 120-case medium test suite and 52/52 on a newly added 52-case issue-regression set, while reducing base prompt length by 55.67% versus `v35`. The project includes a Python deployment script (`codex-instruct.py`) with dry-run, rollback, and snapshot features, plus reproducible evaluation tooling covering 360 cases across 6 scenario types. An optional skills-compatible variant (`v41-skills`) integrates upstream Agent Skills for reverse engineering scenarios. The project is MIT-licensed, actively maintained with CI across Python 3.8/3.13, and is trending likely due to its novel self-improving jailbreak methodology and high benchmark scores against a high-profile model name.